Writing an AI Image Disclosure Customers Can Understand

“AI may have been used” is technically cautious and practically useless. It does not tell a customer whether the background changed, the garment changed, the person is fictional, or the entire image was generated.

A good disclosure answers three questions in plain language:

1. What is synthetic or edited?
2. What should the viewer not treat as evidence?
3. Where can the viewer find more detail if it matters?

Start with the likely misunderstanding

Disclosure is not a ritual phrase. It is a correction to a reasonable interpretation of the image.

If a fictional model wears a fictional dress, the risk is that viewers believe both exist. If a real dress appears on a synthetic model, the risk is that viewers treat the body, drape, or fit as observed. If only the background changed, the risk may be that viewers believe the campaign was photographed at a real location.

Write the disclosure to prevent that specific misunderstanding.

A practical disclosure ladder

Level 1: Minor presentational edit

Use when the product remains unchanged and an element such as background extension or object removal is not material to the buying decision.

Short label: “Background edited with AI.”

Expanded caption: “The product is photographed from the sellable sample. AI was used only to extend the studio background.”

Level 2: Synthetic setting or styling

Use when the product is real but the scene, props, or styling context is generated.

Short label: “AI-generated setting; actual product shown.”

Expanded caption: “The garment is the actual style. The room and props are AI-generated and do not depict a real location.”

Level 3: Synthetic person or fit visualization

Use when the person is generated or a garment is visualized on a person.

Short label: “Fictional AI model; not fit evidence.”

Expanded caption: “The person is AI-generated. The image is a styling visualization and does not demonstrate actual fit, scale, or fabric behavior.”

Level 4: Fully synthetic concept

Use when the person, garment, and setting are generated.

Short label: “AI-generated concept image.”

Expanded caption: “The model, clothing, and location are fictional and AI-generated. No pictured item is offered for sale.”

Level 5: Composite with real and synthetic elements

Short label: “Composite image with AI-generated elements.”

Expanded caption: “The bag is photographed from the actual product. The person, clothing, and street scene are AI-generated.”

Specific nouns do more work than the phrase “AI-assisted.”

Put the information where the impression happens

FTC digital-disclosure guidance emphasizes that a disclosure must be clear and conspicuous and cannot cure a false claim. It also advises evaluating the entire ad, including text and depictions, rather than isolating one phrase.

For a fashion image, that means:

• place the short label immediately below or over the image in a readable area;
• repeat it for each standalone social asset rather than relying on a profile bio;
• keep it visible before a purchase decision, not only in terms and conditions;
• avoid vague icons that require prior knowledge;
• use sufficient contrast and type size;
• keep labels away from edges that common crops may remove;
• preserve the disclosure in carousels, video frames, reposts, and exported variants.

Google’s current Merchant Center guidance similarly warns that self-applied AI labels should not sit at image corners because different placements may crop them. Its platform disclosure can appear in a “How this ad was made” panel and, in some regions, on the ad itself.

Write alt text for the image, not the production log

Alt text should still describe what the image communicates. Add synthetic context when a person using a screen reader would otherwise receive a materially different impression.

Example:

> “AI-generated editorial image of a fictional model in a cobalt coat standing in a concrete studio; the image is a campaign concept, not product-fit evidence.”

Do not stuff the alt text with tool names, prompt history, legal boilerplate, or every edit. Link to a detailed disclosure when the production history is relevant.

Machine-readable provenance is a second channel

Google Merchant Center requires IPTC digital-source metadata for generative-AI product images and structured fields for AI-generated titles and descriptions. C2PA Content Credentials can record origin, edits, and AI use in a cryptographically bound manifest.

Those systems are useful, but C2PA’s own explainer says provenance does not decide whether an image is true. Metadata can also be stripped by screenshots, compression, or unsupported distribution paths. Use machine-readable provenance with visible language, not instead of it.

Words to avoid

“Enhanced.” It does not say whether color correction or a new person was added.

“Digitally created.” It could describe ordinary photography processing.

“For illustration only.” It does not identify which element is fictional.

“AI may have been used.” It sounds uncertain even when the publisher knows the workflow.

A tool name alone. “Made with [tool]” says who supplied software, not what changed.

Test the disclosure

Show the final asset, at the size and duration customers will encounter it, to five people who did not make it. Ask:

1. Which parts do you believe are real?
2. Is the pictured product available to buy?
3. Does the image show how the item actually fits?
4. Where did you notice information about AI?
5. What do you still not understand?

Do not begin by asking “Did you see the disclosure?” That question tells the participant what to look for. Record misunderstandings, revise one variable—wording, position, contrast, or timing—and test again.

Keep a durable record

For each published asset, retain:

• final visible disclosure;
• expanded caption;
• alt text;
• source images and rights;
• prompt and edit history;
• model or tool used;
• embedded provenance status;
• approving editor and date;
• live URLs and social variants;
• correction or removal history.

The reader should not need that full archive. The publisher does.

The one-sentence rule

If a disclosure cannot say what is synthetic in one direct sentence, the production team probably has not defined the image clearly enough to publish it.

Sources and verification

• FTC: .com Disclosures — How to Make Effective Disclosures in Digital Advertising, March 2013.
• FTC: Updated Endorsement Guides, June 29, 2023.
• Google Merchant Center: Use AI content label settings and disclosures, accessed August 31, 2026.
• Google Merchant Center: AI-generated content, accessed August 31, 2026.
• C2PA: Content Credentials Explainer, accessed August 31, 2026.
• NIST: Synthetic Content Transparency Methods, primary government report.